Laser & Optoelectronics Progress, Volume. 61, Issue 10, 1028002(2024)
Remote Sensing Image Semantic Segmentation Algorithm Based on TransMANet
Figures & Tables(20)
Fig. 1. Schematic of MANet. (a) Architecture of MANet; (b) attention block; (c) decoder block
Fig. 6. Schematic of local attention module. (a) Feature pyramid channel; (b) architecture of local attention module
Fig. 9. Dataset presentation. (a) UAVid dataset; (b) LoveDA dataset; (c) Vaihingen dataset; (d) Potsdam dataset
Fig. 10. Visualization results on the UAVid test set. (a) Original image; (b) MANet; (c) TransMANet
Fig. 11. Visualization results on the LoveDA test set. (a) Original image; (b) MANet; (c) TransMANet
Table 1. Performance comparison of different methods on UAVid dataset
View tableTable 1. Performance comparison of different methods on UAVid dataset
Model IoU/% mIoU /% Background Building Road Tree Vegetation Static car Moving car Human Segmenter［ 26 ］64.2 84.4 79.8 76.1 57.6 34.5 59.2 14.2 58.7 SwiftNet［ 24 ］64.1 85.3 61.5 78.3 76.4 62.1 51.1 15.7 61.1 ABCNet［ 25 ］67.4 86.4 81.2 79.9 63.1 48.4 69.8 13.9 63.8 SegFormer［ 27 ］66.6 86.3 80.1 79.6 62.3 52.5 72.5 28.5 66.0 MANet［ 17 ］67.4 88.0 79.5 79.4 63.0 64.6 67.4 21.7 66.3 UNetFormer［ 16 ］68.4 87.4 81.5 80.2 63.5 56.4 73.6 31.0 67.8 TransMANet 69.9 88.7 81.8 80.7 63.6 67.7 73.9 33.2 69.9
Table 2. Performance comparison of different methods on LoveDA dataset
View tableTable 2. Performance comparison of different methods on LoveDA dataset
Model IoU /% mIoU /% Background Building Road Water Barren Forest Agriculture TransUNet［ 23 ］43.0 56.1 53.7 78.0 9.3 44.9 56.9 48.9 Segmenter［ 26 ］38.0 50.7 48.7 77.4 13.3 43.5 58.2 47.1 SwinUperNet［ 20 ］43.3 54.3 54.3 78.7 14.9 45.3 59.6 50.0 DC-Swin［ 28 ］41.3 54.5 56.2 78.1 14.5 47.2 62.4 50.6 MANet［ 17 ］46.5 57.5 53.2 79.3 18.3 47.0 67.2 52.7 UNetFormer［ 16 ］44.7 58.8 54.9 79.6 20.1 46.0 62.5 52.4 TransMANet 46.2 59.7 60.9 82.3 19.7 49.9 65.3 54.8
Table 3. Performance comparison of different methods on Vaihingen dataset
View tableTable 3. Performance comparison of different methods on Vaihingen dataset
Model F1/% Mean F1 /% OA /% mIoU /% Impervious surface Building Low vegetation Tree Car LANet［ 21 ］92.4 94.9 82.9 88.9 81.3 88.1 89.8 DeepLabV3+［ 9 ］91.6 94.1 82.5 88.0 77.7 86.7 89.1 77.1 SwiftNet［ 24 ］92.2 94.8 84.1 89.3 81.2 88.3 90.2 79.6 ABCNet［ 25 ］92.7 95.2 84.5 89.7 85.3 89.5 90.7 81.3 MANet［ 17 ］93.0 95.4 84.6 90.0 88.9 90.4 91.0 82.7 UNetFormer［ 16 ］92.7 95.3 84.9 90.6 88.5 90.4 91.0 82.7 TransMANet 93.4 95.9 85.6 90.7 91.0 91.3 91.5 84.9
Table 4. Performance comparison of different methods on Potsdam dataset
View tableTable 4. Performance comparison of different methods on Potsdam dataset
Model F1 /% Mean F1 /% OA /% mIoU /% Impervious surface Building Low vegetation Tree Car LANet［ 21 ］93.1 97.2 87.3 88.0 94.2 92.0 90.8 DeepLabV3+［ 9 ］92.1 95.3 85.6 86.5 94.8 90.9 89.2 84.2 SwiftNet［ 24 ］91.8 95.9 85.7 86.8 94.5 91.0 89.3 83.8 ABCNet［ 25 ］93.5 96.9 87.9 89.1 95.8 92.7 91.3 86.5 MANet［ 17 ］93.4 97.0 88.3 89.4 96.5 92.9 91.3 87.0 UNetFormer［ 16 ］93.6 97.2 87.7 88.9 96.5 92.8 91.3 86.8 TransMANet 93.7 96.9 88.3 89.6 96.9 93.1 91.7 87.3
Table 5. Comparison of parameter number and calculation amount
View tableTable 5. Comparison of parameter number and calculation amount
Model Parameters /106 FLOPs /109 Segmenter［ 26 ］6.7 26.8 SwiftNet［ 24 ］11.8 51.6 ABCNet［ 25 ］14.0 62.9 SegFormer［ 27 ］13.7 63.3 MANet［ 17 ］35.9 311.6 UNetFormer［ 16 ］11.7 46.9 TransUNet［ 23 ］90.7 803.4 SwinUperNet［ 20 ］60.0 349.1 DC-Swin［ 28 ］45.6 183.8 LANat［ 21 ］23.8 22.0 DeepLabV3+［ 9 ］39.7 30.7 TransMANet 27.2 112.5
Table 6. Ablation study on LoveDA dataset
View tableTable 6. Ablation study on LoveDA dataset
Model IoU /% mIoU /% Background Building Road Water Barren Forest Agriculture Base 46.5 57.6 53.2 79.3 18.3 47.0 67.2 52.7 Base+① 45.7 57.6 54.8 80.0 16.5 47.4 65.8 52.5 Base+② 46.4 60.2 56.4 80.7 18.4 48.9 63.5 53.5 Base+③ 45.5 58.9 57.4 80.2 18.7 48.1 66.5 53.6 Base+①②③ 46.2 59.7 60.9 82.3 19.7 49.9 65.3 54.8
Table 7. Ablation study on UAVid dataset
View tableTable 7. Ablation study on UAVid dataset
Model IoU /% mIoU /% OA /% Background Building Road Tree Vegetation Static car Moving car Human Base 67.4 87.7 79.5 79.4 63.0 64.6 67.4 21.7 66.3 86.1 Base+① 68.8 88.2 81.5 80.2 64.3 64.1 69.7 32.8 68.7 86.9 Base+② 69.6 89.1 81.3 80.7 63.8 63.4 68.7 24.7 67.6 87.3 Base+③ 68.6 89.0 80.8 80.2 63.4 59.1 63.0 23.9 66.0 86.8 Base+①②③ 69.9 88.7 81.7 80.7 63.6 67.7 73.9 33.2 69.9 87.4
Table 8. Ablation study on Vaihingen dataset
View tableTable 8. Ablation study on Vaihingen dataset
Model F1 /% Mean F1 /% OA /% mIoU /% Impervious surface Building Low vegetation Tree Car Base 93.0 95.4 84.6 90.0 88.9 90.4 91.0 82.7 Base+① 93.0 95.8 85.8 90.7 86.9 90.5 91.4 82.8 Base+② 93.4 96.0 85.7 90.7 89.1 91.0 91.6 83.6 Base+③ 93.2 96.0 85.5 90.7 88.7 90.8 91.5 83.4 Base+①②③ 93.4 95.9 85.6 90.7 91.0 91.3 91.5 84.9
Table 9. Ablation study on Potsdam dataset
View tableTable 9. Ablation study on Potsdam dataset
Model F1 /% Mean F1 /% OA /% mIoU /% Impervious surface Building Low vegetation Tree Car Base 93.4 97.0 88.3 89.4 96.5 92.9 91.3 87.0 Base+① 93.1 96.7 88.3 89.6 96.5 92.9 91.3 86.9 Base+② 93.5 97.2 88.0 89.7 96.4 93.0 91.6 87.1 Base+③ 93.6 97.0 88.2 89.5 96.3 92.9 91.6 87.0 Base+①②③ 93.7 96.9 88.3 89.6 96.9 93.1 91.7 87.3
Tools
Get Citation
Copy Citation Text
Xirui Song, Hongwei Ge. Remote Sensing Image Semantic Segmentation Algorithm Based on TransMANet[J]. Laser & Optoelectronics Progress, 2024, 61(10): 1028002
Paper Information
Category: Remote Sensing and Sensors
Received: Sep. 5, 2023
Accepted: Oct. 13, 2023
Published Online: Apr. 29, 2024
The Author Email: Hongwei Ge (ghw8601@163.com)
CSTR:32186.14.LOP232052
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20